[BioC] Unexpectedly high FDR using contrasts in Limma?

[J.delasHeras at ed.ac.uk]

Hi Gordon,

> It all looks normal and expected to me.  Are you surprised that you   
> get a significant MBD1
> comparison but not DBL-MBD1?  Don't forget that direct comparison   
> are more precision, and hence
> more powerful, than indirect comparisons.

I was REALLY hoping that I was doing something wrong... :(

I know that the direct comparison is more powerful (although I don't  
know just *how much* more powerful), and the problem here is that we  
might want to use different controls in the future, so this way I can  
add a new control experiment easily, without repeating the whole  
experiment again.

I am surprised and not. It surprises me because the difference is  
remarkable (at least it looks that way to me) between the DBL  
experiment (no effect) and the MBD1 experiment (clearcut effect), at  
least for 10-20 genes. At the same time, these experiments were tough,  
in that there is a lot of biological variation: for each experiment 3  
replicates were made, and they vary a lot in the level of the observed  
effect... so I guess yeah, it's not that surprising to see that when I  
make a contrast, the overall variation is high enough that it makes it  
too difficult to pick out any individual gene with any degree of  
confidence.
I was still hoping, 'though :-)

> Look at the  stdev.unscaled from limma and you'll see
> that larger values for DBL-MBD1.

Indeed, from 0.58 on MBD1 to 0.87 in DBL-MBD1.

> Other things like including a dye-effect, background correcting, or   
> otherwise checking the data
> often help give more powerful results.

I thought the dye effect is incorporated by default in limma, without  
specifying it implicitly. I noticed that background correction can  
have a big effect on the final results. I really don't like straight  
substraction of background estimates, when the estimate is simply the  
signal measured locally around the spots. I try to use estimates based  
on signal from spots that have DNA spotted but don't hybridise, if I  
can, and substract that. I've tried other methods, normexp can work  
pretty well once you hit the right offset...

I've been meaning to try single channel analysis in limma. I wonder if  
that could help me improve matters a little, by allowing me to compare  
MBD1 against DBL directly... however, the bottomline remains that  
there's a lot of variation in these experiments so if the preliminary  
results look ok, we might just have to increase the number of slides  
(we have the slides, for free, it's only the cost of transfections +  
labelling).

Hmmm, gotta think.

many thanks for your reply, Gordon, it was helpful as usual, just not  
what I wanted to hear :-)

best,

Jose

-- 
Dr. Jose I. de las Heras                      Email: J.delasHeras at ed.ac.uk
The Wellcome Trust Centre for Cell Biology    Phone: +44 (0)131 6513374
Institute for Cell & Molecular Biology        Fax:   +44 (0)131 6507360
Swann Building, Mayfield Road
University of Edinburgh
Edinburgh EH9 3JR
UK